root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 11:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Fix Metal backend broken from the allocator changes (#2455)

Browse Source 
* fix Metal backend broken from the allocator changes
This commit is contained in:
slaren 2023-07-31 11:02:53 +02:00 committed by GitHub
parent a113689571
commit 9d2382b3e4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 10 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
llama.cpp
View File

@ -1812,6 +1812,12 @@ static bool llama_eval_internal(
// otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads; n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
LLAMA_ASSERT(strcmp(res->name, "result_output") == 0);
LLAMA_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
#if GGML_USE_MPI #if GGML_USE_MPI
const int64_t n_layer = hparams.n_layer; const int64_t n_layer = hparams.n_layer;
ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer); ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
@ -1825,7 +1831,10 @@ static bool llama_eval_internal(
//} //}
ggml_metal_set_n_cb (lctx.ctx_metal, n_threads); ggml_metal_set_n_cb (lctx.ctx_metal, n_threads);
ggml_metal_graph_compute(lctx.ctx_metal, gf); ggml_metal_graph_compute(lctx.ctx_metal, gf);
ggml_metal_get_tensor (lctx.ctx_metal, cur); ggml_metal_get_tensor (lctx.ctx_metal, res);
if (!lctx.embedding.empty()) {
ggml_metal_get_tensor(lctx.ctx_metal, embeddings);
}
} else { } else {
// IMPORTANT: // IMPORTANT:
// Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla // Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla
@ -1856,12 +1865,6 @@ static bool llama_eval_internal(
// update kv token count // update kv token count
lctx.kv_self.n = n_past + N; lctx.kv_self.n = n_past + N;
struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
LLAMA_ASSERT(strcmp(res->name, "result_output") == 0);
LLAMA_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
if (cgraph_fname) { if (cgraph_fname) {
ggml_graph_export(gf, cgraph_fname); ggml_graph_export(gf, cgraph_fname);
} }